1//==- SystemZInstrFP.td - Floating-point SystemZ instructions --*- tblgen-*-==// 2// 3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4// See https://llvm.org/LICENSE.txt for license information. 5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6// 7//===----------------------------------------------------------------------===// 8 9// TODO: Most floating-point instructions (except for simple moves and the 10// like) can raise exceptions -- should they have hasSideEffects=1 ? 11 12//===----------------------------------------------------------------------===// 13// Select instructions 14//===----------------------------------------------------------------------===// 15 16// C's ?: operator for floating-point operands. 17let Predicates = [FeatureVector] in { 18 def SelectVR32 : SelectWrapper<f32, VR32>; 19 def SelectVR64 : SelectWrapper<f64, VR64>; 20} 21def SelectF32 : SelectWrapper<f32, FP32>; 22def SelectF64 : SelectWrapper<f64, FP64>; 23let Predicates = [FeatureNoVectorEnhancements1] in 24 def SelectF128 : SelectWrapper<f128, FP128>; 25let Predicates = [FeatureVectorEnhancements1] in 26 def SelectVR128 : SelectWrapper<f128, VR128>; 27 28defm CondStoreF32 : CondStores<FP32, simple_store, 29 simple_load, bdxaddr20only>; 30defm CondStoreF64 : CondStores<FP64, simple_store, 31 simple_load, bdxaddr20only>; 32 33//===----------------------------------------------------------------------===// 34// Move instructions 35//===----------------------------------------------------------------------===// 36 37// Load zero. 38let isAsCheapAsAMove = 1, isMoveImm = 1 in { 39 def LZER : InherentRRE<"lzer", 0xB374, FP32, fpimm0>; 40 def LZDR : InherentRRE<"lzdr", 0xB375, FP64, fpimm0>; 41 def LZXR : InherentRRE<"lzxr", 0xB376, FP128, fpimm0>; 42} 43 44// Moves between two floating-point registers. 45def LER : UnaryRR <"ler", 0x38, null_frag, FP32, FP32>; 46def LDR : UnaryRR <"ldr", 0x28, null_frag, FP64, FP64>; 47def LXR : UnaryRRE<"lxr", 0xB365, null_frag, FP128, FP128>; 48 49// For z13 we prefer LDR over LER to avoid partial register dependencies. 50let isCodeGenOnly = 1 in 51 def LDR32 : UnaryRR<"ldr", 0x28, null_frag, FP32, FP32>; 52 53// Moves between two floating-point registers that also set the condition 54// codes. 55let Uses = [FPC], mayRaiseFPException = 1, 56 Defs = [CC], CCValues = 0xF, CompareZeroCCMask = 0xF in { 57 defm LTEBR : LoadAndTestRRE<"ltebr", 0xB302, FP32>; 58 defm LTDBR : LoadAndTestRRE<"ltdbr", 0xB312, FP64>; 59 defm LTXBR : LoadAndTestRRE<"ltxbr", 0xB342, FP128>; 60} 61// Note that LTxBRCompare is not available if we have vector support, 62// since load-and-test instructions will partially clobber the target 63// (vector) register. 64let Predicates = [FeatureNoVector] in { 65 defm : CompareZeroFP<LTEBRCompare, FP32>; 66 defm : CompareZeroFP<LTDBRCompare, FP64>; 67 defm : CompareZeroFP<LTXBRCompare, FP128>; 68} 69 70// Use a normal load-and-test for compare against zero in case of 71// vector support (via a pseudo to simplify instruction selection). 72let Uses = [FPC], mayRaiseFPException = 1, 73 Defs = [CC], usesCustomInserter = 1, hasNoSchedulingInfo = 1 in { 74 def LTEBRCompare_VecPseudo : Pseudo<(outs), (ins FP32:$R1, FP32:$R2), []>; 75 def LTDBRCompare_VecPseudo : Pseudo<(outs), (ins FP64:$R1, FP64:$R2), []>; 76 def LTXBRCompare_VecPseudo : Pseudo<(outs), (ins FP128:$R1, FP128:$R2), []>; 77} 78let Predicates = [FeatureVector] in { 79 defm : CompareZeroFP<LTEBRCompare_VecPseudo, FP32>; 80 defm : CompareZeroFP<LTDBRCompare_VecPseudo, FP64>; 81} 82let Predicates = [FeatureVector, FeatureNoVectorEnhancements1] in 83 defm : CompareZeroFP<LTXBRCompare_VecPseudo, FP128>; 84 85// Moves between 64-bit integer and floating-point registers. 86def LGDR : UnaryRRE<"lgdr", 0xB3CD, bitconvert, GR64, FP64>; 87def LDGR : UnaryRRE<"ldgr", 0xB3C1, bitconvert, FP64, GR64>; 88 89// fcopysign with an FP32 result. 90let isCodeGenOnly = 1 in { 91 def CPSDRss : BinaryRRFb<"cpsdr", 0xB372, fcopysign, FP32, FP32, FP32>; 92 def CPSDRsd : BinaryRRFb<"cpsdr", 0xB372, fcopysign, FP32, FP32, FP64>; 93} 94 95// The sign of an FP128 is in the high register. 96let Predicates = [FeatureNoVectorEnhancements1] in 97 def : Pat<(fcopysign FP32:$src1, (f32 (fpround (f128 FP128:$src2)))), 98 (CPSDRsd FP32:$src1, (EXTRACT_SUBREG FP128:$src2, subreg_h64))>; 99let Predicates = [FeatureVectorEnhancements1] in 100 def : Pat<(fcopysign FP32:$src1, (f32 (fpround (f128 VR128:$src2)))), 101 (CPSDRsd FP32:$src1, (EXTRACT_SUBREG VR128:$src2, subreg_h64))>; 102 103// fcopysign with an FP64 result. 104let isCodeGenOnly = 1 in 105 def CPSDRds : BinaryRRFb<"cpsdr", 0xB372, fcopysign, FP64, FP64, FP32>; 106def CPSDRdd : BinaryRRFb<"cpsdr", 0xB372, fcopysign, FP64, FP64, FP64>; 107 108// The sign of an FP128 is in the high register. 109let Predicates = [FeatureNoVectorEnhancements1] in 110 def : Pat<(fcopysign FP64:$src1, (f64 (fpround (f128 FP128:$src2)))), 111 (CPSDRdd FP64:$src1, (EXTRACT_SUBREG FP128:$src2, subreg_h64))>; 112let Predicates = [FeatureVectorEnhancements1] in 113 def : Pat<(fcopysign FP64:$src1, (f64 (fpround (f128 VR128:$src2)))), 114 (CPSDRdd FP64:$src1, (EXTRACT_SUBREG VR128:$src2, subreg_h64))>; 115 116// fcopysign with an FP128 result. Use "upper" as the high half and leave 117// the low half as-is. 118class CopySign128<RegisterOperand cls, dag upper> 119 : Pat<(fcopysign FP128:$src1, cls:$src2), 120 (INSERT_SUBREG FP128:$src1, upper, subreg_h64)>; 121 122let Predicates = [FeatureNoVectorEnhancements1] in { 123 def : CopySign128<FP32, (CPSDRds (EXTRACT_SUBREG FP128:$src1, subreg_h64), 124 FP32:$src2)>; 125 def : CopySign128<FP64, (CPSDRdd (EXTRACT_SUBREG FP128:$src1, subreg_h64), 126 FP64:$src2)>; 127 def : CopySign128<FP128, (CPSDRdd (EXTRACT_SUBREG FP128:$src1, subreg_h64), 128 (EXTRACT_SUBREG FP128:$src2, subreg_h64))>; 129} 130 131// The length is given as one less for MVCImm. 132defm LoadStoreF32 : MVCLoadStore<load, f32, MVCImm, 3>; 133defm LoadStoreF64 : MVCLoadStore<load, f64, MVCImm, 7>; 134defm LoadStoreF128 : MVCLoadStore<load, f128, MVCImm, 15>; 135 136//===----------------------------------------------------------------------===// 137// Load instructions 138//===----------------------------------------------------------------------===// 139 140let canFoldAsLoad = 1, SimpleBDXLoad = 1, mayLoad = 1 in { 141 defm LE : UnaryRXPair<"le", 0x78, 0xED64, load, FP32, 4>; 142 defm LD : UnaryRXPair<"ld", 0x68, 0xED65, load, FP64, 8>; 143 144 // For z13 we prefer LDE over LE to avoid partial register dependencies. 145 let isCodeGenOnly = 1 in 146 def LDE32 : UnaryRXE<"lde", 0xED24, null_frag, FP32, 4>; 147 148 // These instructions are split after register allocation, so we don't 149 // want a custom inserter. 150 let Has20BitOffset = 1, HasIndex = 1, Is128Bit = 1 in { 151 def LX : Pseudo<(outs FP128:$dst), (ins bdxaddr20only128:$src), 152 [(set FP128:$dst, (load bdxaddr20only128:$src))]>; 153 } 154} 155 156//===----------------------------------------------------------------------===// 157// Store instructions 158//===----------------------------------------------------------------------===// 159 160let SimpleBDXStore = 1, mayStore = 1 in { 161 defm STE : StoreRXPair<"ste", 0x70, 0xED66, store, FP32, 4>; 162 defm STD : StoreRXPair<"std", 0x60, 0xED67, store, FP64, 8>; 163 164 // These instructions are split after register allocation, so we don't 165 // want a custom inserter. 166 let Has20BitOffset = 1, HasIndex = 1, Is128Bit = 1 in { 167 def STX : Pseudo<(outs), (ins FP128:$src, bdxaddr20only128:$dst), 168 [(store FP128:$src, bdxaddr20only128:$dst)]>; 169 } 170} 171 172//===----------------------------------------------------------------------===// 173// Conversion instructions 174//===----------------------------------------------------------------------===// 175 176// Convert floating-point values to narrower representations, rounding 177// according to the current mode. The destination of LEXBR and LDXBR 178// is a 128-bit value, but only the first register of the pair is used. 179let Uses = [FPC], mayRaiseFPException = 1 in { 180 def LEDBR : UnaryRRE<"ledbr", 0xB344, any_fpround, FP32, FP64>; 181 def LEXBR : UnaryRRE<"lexbr", 0xB346, null_frag, FP128, FP128>; 182 def LDXBR : UnaryRRE<"ldxbr", 0xB345, null_frag, FP128, FP128>; 183 184 def LEDBRA : TernaryRRFe<"ledbra", 0xB344, FP32, FP64>, 185 Requires<[FeatureFPExtension]>; 186 def LEXBRA : TernaryRRFe<"lexbra", 0xB346, FP128, FP128>, 187 Requires<[FeatureFPExtension]>; 188 def LDXBRA : TernaryRRFe<"ldxbra", 0xB345, FP128, FP128>, 189 Requires<[FeatureFPExtension]>; 190} 191 192let Predicates = [FeatureNoVectorEnhancements1] in { 193 def : Pat<(f32 (any_fpround FP128:$src)), 194 (EXTRACT_SUBREG (LEXBR FP128:$src), subreg_h32)>; 195 def : Pat<(f64 (any_fpround FP128:$src)), 196 (EXTRACT_SUBREG (LDXBR FP128:$src), subreg_h64)>; 197} 198 199// Extend register floating-point values to wider representations. 200let Uses = [FPC], mayRaiseFPException = 1 in { 201 def LDEBR : UnaryRRE<"ldebr", 0xB304, any_fpextend, FP64, FP32>; 202 def LXEBR : UnaryRRE<"lxebr", 0xB306, null_frag, FP128, FP32>; 203 def LXDBR : UnaryRRE<"lxdbr", 0xB305, null_frag, FP128, FP64>; 204} 205let Predicates = [FeatureNoVectorEnhancements1] in { 206 def : Pat<(f128 (any_fpextend (f32 FP32:$src))), (LXEBR FP32:$src)>; 207 def : Pat<(f128 (any_fpextend (f64 FP64:$src))), (LXDBR FP64:$src)>; 208} 209 210// Extend memory floating-point values to wider representations. 211let Uses = [FPC], mayRaiseFPException = 1 in { 212 def LDEB : UnaryRXE<"ldeb", 0xED04, any_extloadf32, FP64, 4>; 213 def LXEB : UnaryRXE<"lxeb", 0xED06, null_frag, FP128, 4>; 214 def LXDB : UnaryRXE<"lxdb", 0xED05, null_frag, FP128, 8>; 215} 216let Predicates = [FeatureNoVectorEnhancements1] in { 217 def : Pat<(f128 (any_extloadf32 bdxaddr12only:$src)), 218 (LXEB bdxaddr12only:$src)>; 219 def : Pat<(f128 (any_extloadf64 bdxaddr12only:$src)), 220 (LXDB bdxaddr12only:$src)>; 221} 222 223// Convert a signed integer register value to a floating-point one. 224let Uses = [FPC], mayRaiseFPException = 1 in { 225 def CEFBR : UnaryRRE<"cefbr", 0xB394, any_sint_to_fp, FP32, GR32>; 226 def CDFBR : UnaryRRE<"cdfbr", 0xB395, any_sint_to_fp, FP64, GR32>; 227 def CXFBR : UnaryRRE<"cxfbr", 0xB396, any_sint_to_fp, FP128, GR32>; 228 229 def CEGBR : UnaryRRE<"cegbr", 0xB3A4, any_sint_to_fp, FP32, GR64>; 230 def CDGBR : UnaryRRE<"cdgbr", 0xB3A5, any_sint_to_fp, FP64, GR64>; 231 def CXGBR : UnaryRRE<"cxgbr", 0xB3A6, any_sint_to_fp, FP128, GR64>; 232} 233 234// The FP extension feature provides versions of the above that allow 235// specifying rounding mode and inexact-exception suppression flags. 236let Uses = [FPC], mayRaiseFPException = 1, Predicates = [FeatureFPExtension] in { 237 def CEFBRA : TernaryRRFe<"cefbra", 0xB394, FP32, GR32>; 238 def CDFBRA : TernaryRRFe<"cdfbra", 0xB395, FP64, GR32>; 239 def CXFBRA : TernaryRRFe<"cxfbra", 0xB396, FP128, GR32>; 240 241 def CEGBRA : TernaryRRFe<"cegbra", 0xB3A4, FP32, GR64>; 242 def CDGBRA : TernaryRRFe<"cdgbra", 0xB3A5, FP64, GR64>; 243 def CXGBRA : TernaryRRFe<"cxgbra", 0xB3A6, FP128, GR64>; 244} 245 246// Convert am unsigned integer register value to a floating-point one. 247let Predicates = [FeatureFPExtension] in { 248 let Uses = [FPC], mayRaiseFPException = 1 in { 249 def CELFBR : TernaryRRFe<"celfbr", 0xB390, FP32, GR32>; 250 def CDLFBR : TernaryRRFe<"cdlfbr", 0xB391, FP64, GR32>; 251 def CXLFBR : TernaryRRFe<"cxlfbr", 0xB392, FP128, GR32>; 252 253 def CELGBR : TernaryRRFe<"celgbr", 0xB3A0, FP32, GR64>; 254 def CDLGBR : TernaryRRFe<"cdlgbr", 0xB3A1, FP64, GR64>; 255 def CXLGBR : TernaryRRFe<"cxlgbr", 0xB3A2, FP128, GR64>; 256 } 257 258 def : Pat<(f32 (any_uint_to_fp GR32:$src)), (CELFBR 0, GR32:$src, 0)>; 259 def : Pat<(f64 (any_uint_to_fp GR32:$src)), (CDLFBR 0, GR32:$src, 0)>; 260 def : Pat<(f128 (any_uint_to_fp GR32:$src)), (CXLFBR 0, GR32:$src, 0)>; 261 262 def : Pat<(f32 (any_uint_to_fp GR64:$src)), (CELGBR 0, GR64:$src, 0)>; 263 def : Pat<(f64 (any_uint_to_fp GR64:$src)), (CDLGBR 0, GR64:$src, 0)>; 264 def : Pat<(f128 (any_uint_to_fp GR64:$src)), (CXLGBR 0, GR64:$src, 0)>; 265} 266 267// Convert a floating-point register value to a signed integer value, 268// with the second operand (modifier M3) specifying the rounding mode. 269let Uses = [FPC], mayRaiseFPException = 1, Defs = [CC] in { 270 def CFEBR : BinaryRRFe<"cfebr", 0xB398, GR32, FP32>; 271 def CFDBR : BinaryRRFe<"cfdbr", 0xB399, GR32, FP64>; 272 def CFXBR : BinaryRRFe<"cfxbr", 0xB39A, GR32, FP128>; 273 274 def CGEBR : BinaryRRFe<"cgebr", 0xB3A8, GR64, FP32>; 275 def CGDBR : BinaryRRFe<"cgdbr", 0xB3A9, GR64, FP64>; 276 def CGXBR : BinaryRRFe<"cgxbr", 0xB3AA, GR64, FP128>; 277} 278 279// fp_to_sint always rounds towards zero, which is modifier value 5. 280def : Pat<(i32 (any_fp_to_sint FP32:$src)), (CFEBR 5, FP32:$src)>; 281def : Pat<(i32 (any_fp_to_sint FP64:$src)), (CFDBR 5, FP64:$src)>; 282def : Pat<(i32 (any_fp_to_sint FP128:$src)), (CFXBR 5, FP128:$src)>; 283 284def : Pat<(i64 (any_fp_to_sint FP32:$src)), (CGEBR 5, FP32:$src)>; 285def : Pat<(i64 (any_fp_to_sint FP64:$src)), (CGDBR 5, FP64:$src)>; 286def : Pat<(i64 (any_fp_to_sint FP128:$src)), (CGXBR 5, FP128:$src)>; 287 288// The FP extension feature provides versions of the above that allow 289// also specifying the inexact-exception suppression flag. 290let Uses = [FPC], mayRaiseFPException = 1, 291 Predicates = [FeatureFPExtension], Defs = [CC] in { 292 def CFEBRA : TernaryRRFe<"cfebra", 0xB398, GR32, FP32>; 293 def CFDBRA : TernaryRRFe<"cfdbra", 0xB399, GR32, FP64>; 294 def CFXBRA : TernaryRRFe<"cfxbra", 0xB39A, GR32, FP128>; 295 296 def CGEBRA : TernaryRRFe<"cgebra", 0xB3A8, GR64, FP32>; 297 def CGDBRA : TernaryRRFe<"cgdbra", 0xB3A9, GR64, FP64>; 298 def CGXBRA : TernaryRRFe<"cgxbra", 0xB3AA, GR64, FP128>; 299} 300 301// Convert a floating-point register value to an unsigned integer value. 302let Predicates = [FeatureFPExtension] in { 303 let Uses = [FPC], mayRaiseFPException = 1, Defs = [CC] in { 304 def CLFEBR : TernaryRRFe<"clfebr", 0xB39C, GR32, FP32>; 305 def CLFDBR : TernaryRRFe<"clfdbr", 0xB39D, GR32, FP64>; 306 def CLFXBR : TernaryRRFe<"clfxbr", 0xB39E, GR32, FP128>; 307 308 def CLGEBR : TernaryRRFe<"clgebr", 0xB3AC, GR64, FP32>; 309 def CLGDBR : TernaryRRFe<"clgdbr", 0xB3AD, GR64, FP64>; 310 def CLGXBR : TernaryRRFe<"clgxbr", 0xB3AE, GR64, FP128>; 311 } 312 313 def : Pat<(i32 (any_fp_to_uint FP32:$src)), (CLFEBR 5, FP32:$src, 0)>; 314 def : Pat<(i32 (any_fp_to_uint FP64:$src)), (CLFDBR 5, FP64:$src, 0)>; 315 def : Pat<(i32 (any_fp_to_uint FP128:$src)), (CLFXBR 5, FP128:$src, 0)>; 316 317 def : Pat<(i64 (any_fp_to_uint FP32:$src)), (CLGEBR 5, FP32:$src, 0)>; 318 def : Pat<(i64 (any_fp_to_uint FP64:$src)), (CLGDBR 5, FP64:$src, 0)>; 319 def : Pat<(i64 (any_fp_to_uint FP128:$src)), (CLGXBR 5, FP128:$src, 0)>; 320} 321 322 323//===----------------------------------------------------------------------===// 324// Unary arithmetic 325//===----------------------------------------------------------------------===// 326 327// We prefer generic instructions during isel, because they do not 328// clobber CC and therefore give the scheduler more freedom. In cases 329// the CC is actually useful, the SystemZElimCompare pass will try to 330// convert generic instructions into opcodes that also set CC. Note 331// that lcdf / lpdf / lndf only affect the sign bit, and can therefore 332// be used with fp32 as well. This could be done for fp128, in which 333// case the operands would have to be tied. 334 335// Negation (Load Complement). 336let Defs = [CC], CCValues = 0xF, CompareZeroCCMask = 0xF in { 337 def LCEBR : UnaryRRE<"lcebr", 0xB303, null_frag, FP32, FP32>; 338 def LCDBR : UnaryRRE<"lcdbr", 0xB313, null_frag, FP64, FP64>; 339 def LCXBR : UnaryRRE<"lcxbr", 0xB343, fneg, FP128, FP128>; 340} 341// Generic form, which does not set CC. 342def LCDFR : UnaryRRE<"lcdfr", 0xB373, fneg, FP64, FP64>; 343let isCodeGenOnly = 1 in 344 def LCDFR_32 : UnaryRRE<"lcdfr", 0xB373, fneg, FP32, FP32>; 345 346// Absolute value (Load Positive). 347let Defs = [CC], CCValues = 0xF, CompareZeroCCMask = 0xF in { 348 def LPEBR : UnaryRRE<"lpebr", 0xB300, null_frag, FP32, FP32>; 349 def LPDBR : UnaryRRE<"lpdbr", 0xB310, null_frag, FP64, FP64>; 350 def LPXBR : UnaryRRE<"lpxbr", 0xB340, fabs, FP128, FP128>; 351} 352// Generic form, which does not set CC. 353def LPDFR : UnaryRRE<"lpdfr", 0xB370, fabs, FP64, FP64>; 354let isCodeGenOnly = 1 in 355 def LPDFR_32 : UnaryRRE<"lpdfr", 0xB370, fabs, FP32, FP32>; 356 357// Negative absolute value (Load Negative). 358let Defs = [CC], CCValues = 0xF, CompareZeroCCMask = 0xF in { 359 def LNEBR : UnaryRRE<"lnebr", 0xB301, null_frag, FP32, FP32>; 360 def LNDBR : UnaryRRE<"lndbr", 0xB311, null_frag, FP64, FP64>; 361 def LNXBR : UnaryRRE<"lnxbr", 0xB341, fnabs, FP128, FP128>; 362} 363// Generic form, which does not set CC. 364def LNDFR : UnaryRRE<"lndfr", 0xB371, fnabs, FP64, FP64>; 365let isCodeGenOnly = 1 in 366 def LNDFR_32 : UnaryRRE<"lndfr", 0xB371, fnabs, FP32, FP32>; 367 368// Square root. 369let Uses = [FPC], mayRaiseFPException = 1 in { 370 def SQEBR : UnaryRRE<"sqebr", 0xB314, any_fsqrt, FP32, FP32>; 371 def SQDBR : UnaryRRE<"sqdbr", 0xB315, any_fsqrt, FP64, FP64>; 372 def SQXBR : UnaryRRE<"sqxbr", 0xB316, any_fsqrt, FP128, FP128>; 373 374 def SQEB : UnaryRXE<"sqeb", 0xED14, loadu<any_fsqrt>, FP32, 4>; 375 def SQDB : UnaryRXE<"sqdb", 0xED15, loadu<any_fsqrt>, FP64, 8>; 376} 377 378// Round to an integer, with the second operand (modifier M3) specifying 379// the rounding mode. These forms always check for inexact conditions. 380let Uses = [FPC], mayRaiseFPException = 1 in { 381 def FIEBR : BinaryRRFe<"fiebr", 0xB357, FP32, FP32>; 382 def FIDBR : BinaryRRFe<"fidbr", 0xB35F, FP64, FP64>; 383 def FIXBR : BinaryRRFe<"fixbr", 0xB347, FP128, FP128>; 384} 385 386// frint rounds according to the current mode (modifier 0) and detects 387// inexact conditions. 388def : Pat<(any_frint FP32:$src), (FIEBR 0, FP32:$src)>; 389def : Pat<(any_frint FP64:$src), (FIDBR 0, FP64:$src)>; 390def : Pat<(any_frint FP128:$src), (FIXBR 0, FP128:$src)>; 391 392let Predicates = [FeatureFPExtension] in { 393 // Extended forms of the FIxBR instructions. M4 can be set to 4 394 // to suppress detection of inexact conditions. 395 let Uses = [FPC], mayRaiseFPException = 1 in { 396 def FIEBRA : TernaryRRFe<"fiebra", 0xB357, FP32, FP32>; 397 def FIDBRA : TernaryRRFe<"fidbra", 0xB35F, FP64, FP64>; 398 def FIXBRA : TernaryRRFe<"fixbra", 0xB347, FP128, FP128>; 399 } 400 401 // fnearbyint is like frint but does not detect inexact conditions. 402 def : Pat<(any_fnearbyint FP32:$src), (FIEBRA 0, FP32:$src, 4)>; 403 def : Pat<(any_fnearbyint FP64:$src), (FIDBRA 0, FP64:$src, 4)>; 404 def : Pat<(any_fnearbyint FP128:$src), (FIXBRA 0, FP128:$src, 4)>; 405 406 // floor is no longer allowed to raise an inexact condition, 407 // so restrict it to the cases where the condition can be suppressed. 408 // Mode 7 is round towards -inf. 409 def : Pat<(any_ffloor FP32:$src), (FIEBRA 7, FP32:$src, 4)>; 410 def : Pat<(any_ffloor FP64:$src), (FIDBRA 7, FP64:$src, 4)>; 411 def : Pat<(any_ffloor FP128:$src), (FIXBRA 7, FP128:$src, 4)>; 412 413 // Same idea for ceil, where mode 6 is round towards +inf. 414 def : Pat<(any_fceil FP32:$src), (FIEBRA 6, FP32:$src, 4)>; 415 def : Pat<(any_fceil FP64:$src), (FIDBRA 6, FP64:$src, 4)>; 416 def : Pat<(any_fceil FP128:$src), (FIXBRA 6, FP128:$src, 4)>; 417 418 // Same idea for trunc, where mode 5 is round towards zero. 419 def : Pat<(any_ftrunc FP32:$src), (FIEBRA 5, FP32:$src, 4)>; 420 def : Pat<(any_ftrunc FP64:$src), (FIDBRA 5, FP64:$src, 4)>; 421 def : Pat<(any_ftrunc FP128:$src), (FIXBRA 5, FP128:$src, 4)>; 422 423 // Same idea for round, where mode 1 is round towards nearest with 424 // ties away from zero. 425 def : Pat<(any_fround FP32:$src), (FIEBRA 1, FP32:$src, 4)>; 426 def : Pat<(any_fround FP64:$src), (FIDBRA 1, FP64:$src, 4)>; 427 def : Pat<(any_fround FP128:$src), (FIXBRA 1, FP128:$src, 4)>; 428} 429 430//===----------------------------------------------------------------------===// 431// Binary arithmetic 432//===----------------------------------------------------------------------===// 433 434// Addition. 435let Uses = [FPC], mayRaiseFPException = 1, 436 Defs = [CC], CCValues = 0xF, CompareZeroCCMask = 0xF in { 437 let isCommutable = 1 in { 438 def AEBR : BinaryRRE<"aebr", 0xB30A, any_fadd, FP32, FP32>; 439 def ADBR : BinaryRRE<"adbr", 0xB31A, any_fadd, FP64, FP64>; 440 def AXBR : BinaryRRE<"axbr", 0xB34A, any_fadd, FP128, FP128>; 441 } 442 defm AEB : BinaryRXEAndPseudo<"aeb", 0xED0A, any_fadd, FP32, load, 4>; 443 defm ADB : BinaryRXEAndPseudo<"adb", 0xED1A, any_fadd, FP64, load, 8>; 444} 445 446// Subtraction. 447let Uses = [FPC], mayRaiseFPException = 1, 448 Defs = [CC], CCValues = 0xF, CompareZeroCCMask = 0xF in { 449 def SEBR : BinaryRRE<"sebr", 0xB30B, any_fsub, FP32, FP32>; 450 def SDBR : BinaryRRE<"sdbr", 0xB31B, any_fsub, FP64, FP64>; 451 def SXBR : BinaryRRE<"sxbr", 0xB34B, any_fsub, FP128, FP128>; 452 453 defm SEB : BinaryRXEAndPseudo<"seb", 0xED0B, any_fsub, FP32, load, 4>; 454 defm SDB : BinaryRXEAndPseudo<"sdb", 0xED1B, any_fsub, FP64, load, 8>; 455} 456 457// Multiplication. 458let Uses = [FPC], mayRaiseFPException = 1 in { 459 let isCommutable = 1 in { 460 def MEEBR : BinaryRRE<"meebr", 0xB317, any_fmul, FP32, FP32>; 461 def MDBR : BinaryRRE<"mdbr", 0xB31C, any_fmul, FP64, FP64>; 462 def MXBR : BinaryRRE<"mxbr", 0xB34C, any_fmul, FP128, FP128>; 463 } 464 defm MEEB : BinaryRXEAndPseudo<"meeb", 0xED17, any_fmul, FP32, load, 4>; 465 defm MDB : BinaryRXEAndPseudo<"mdb", 0xED1C, any_fmul, FP64, load, 8>; 466} 467 468// f64 multiplication of two FP32 registers. 469let Uses = [FPC], mayRaiseFPException = 1 in 470 def MDEBR : BinaryRRE<"mdebr", 0xB30C, null_frag, FP64, FP32>; 471def : Pat<(any_fmul (f64 (any_fpextend FP32:$src1)), 472 (f64 (any_fpextend FP32:$src2))), 473 (MDEBR (INSERT_SUBREG (f64 (IMPLICIT_DEF)), 474 FP32:$src1, subreg_h32), FP32:$src2)>; 475 476// f64 multiplication of an FP32 register and an f32 memory. 477let Uses = [FPC], mayRaiseFPException = 1 in 478 def MDEB : BinaryRXE<"mdeb", 0xED0C, null_frag, FP64, load, 4>; 479def : Pat<(any_fmul (f64 (any_fpextend FP32:$src1)), 480 (f64 (any_extloadf32 bdxaddr12only:$addr))), 481 (MDEB (INSERT_SUBREG (f64 (IMPLICIT_DEF)), FP32:$src1, subreg_h32), 482 bdxaddr12only:$addr)>; 483 484// f128 multiplication of two FP64 registers. 485let Uses = [FPC], mayRaiseFPException = 1 in 486 def MXDBR : BinaryRRE<"mxdbr", 0xB307, null_frag, FP128, FP64>; 487let Predicates = [FeatureNoVectorEnhancements1] in 488 def : Pat<(any_fmul (f128 (any_fpextend FP64:$src1)), 489 (f128 (any_fpextend FP64:$src2))), 490 (MXDBR (INSERT_SUBREG (f128 (IMPLICIT_DEF)), 491 FP64:$src1, subreg_h64), FP64:$src2)>; 492 493// f128 multiplication of an FP64 register and an f64 memory. 494let Uses = [FPC], mayRaiseFPException = 1 in 495 def MXDB : BinaryRXE<"mxdb", 0xED07, null_frag, FP128, load, 8>; 496let Predicates = [FeatureNoVectorEnhancements1] in 497 def : Pat<(any_fmul (f128 (any_fpextend FP64:$src1)), 498 (f128 (any_extloadf64 bdxaddr12only:$addr))), 499 (MXDB (INSERT_SUBREG (f128 (IMPLICIT_DEF)), FP64:$src1, subreg_h64), 500 bdxaddr12only:$addr)>; 501 502// Fused multiply-add. 503let Uses = [FPC], mayRaiseFPException = 1 in { 504 def MAEBR : TernaryRRD<"maebr", 0xB30E, z_any_fma, FP32, FP32>; 505 def MADBR : TernaryRRD<"madbr", 0xB31E, z_any_fma, FP64, FP64>; 506 507 defm MAEB : TernaryRXFAndPseudo<"maeb", 0xED0E, z_any_fma, FP32, FP32, load, 4>; 508 defm MADB : TernaryRXFAndPseudo<"madb", 0xED1E, z_any_fma, FP64, FP64, load, 8>; 509} 510 511// Fused multiply-subtract. 512let Uses = [FPC], mayRaiseFPException = 1 in { 513 def MSEBR : TernaryRRD<"msebr", 0xB30F, z_any_fms, FP32, FP32>; 514 def MSDBR : TernaryRRD<"msdbr", 0xB31F, z_any_fms, FP64, FP64>; 515 516 defm MSEB : TernaryRXFAndPseudo<"mseb", 0xED0F, z_any_fms, FP32, FP32, load, 4>; 517 defm MSDB : TernaryRXFAndPseudo<"msdb", 0xED1F, z_any_fms, FP64, FP64, load, 8>; 518} 519 520// Division. 521let Uses = [FPC], mayRaiseFPException = 1 in { 522 def DEBR : BinaryRRE<"debr", 0xB30D, any_fdiv, FP32, FP32>; 523 def DDBR : BinaryRRE<"ddbr", 0xB31D, any_fdiv, FP64, FP64>; 524 def DXBR : BinaryRRE<"dxbr", 0xB34D, any_fdiv, FP128, FP128>; 525 526 defm DEB : BinaryRXEAndPseudo<"deb", 0xED0D, any_fdiv, FP32, load, 4>; 527 defm DDB : BinaryRXEAndPseudo<"ddb", 0xED1D, any_fdiv, FP64, load, 8>; 528} 529 530// Divide to integer. 531let Uses = [FPC], mayRaiseFPException = 1, Defs = [CC] in { 532 def DIEBR : TernaryRRFb<"diebr", 0xB353, FP32, FP32, FP32>; 533 def DIDBR : TernaryRRFb<"didbr", 0xB35B, FP64, FP64, FP64>; 534} 535 536//===----------------------------------------------------------------------===// 537// Comparisons 538//===----------------------------------------------------------------------===// 539 540let Uses = [FPC], mayRaiseFPException = 1, Defs = [CC], CCValues = 0xF in { 541 def CEBR : CompareRRE<"cebr", 0xB309, z_any_fcmp, FP32, FP32>; 542 def CDBR : CompareRRE<"cdbr", 0xB319, z_any_fcmp, FP64, FP64>; 543 def CXBR : CompareRRE<"cxbr", 0xB349, z_any_fcmp, FP128, FP128>; 544 545 def CEB : CompareRXE<"ceb", 0xED09, z_any_fcmp, FP32, load, 4>; 546 def CDB : CompareRXE<"cdb", 0xED19, z_any_fcmp, FP64, load, 8>; 547 548 def KEBR : CompareRRE<"kebr", 0xB308, z_strict_fcmps, FP32, FP32>; 549 def KDBR : CompareRRE<"kdbr", 0xB318, z_strict_fcmps, FP64, FP64>; 550 def KXBR : CompareRRE<"kxbr", 0xB348, z_strict_fcmps, FP128, FP128>; 551 552 def KEB : CompareRXE<"keb", 0xED08, z_strict_fcmps, FP32, load, 4>; 553 def KDB : CompareRXE<"kdb", 0xED18, z_strict_fcmps, FP64, load, 8>; 554} 555 556// Test Data Class. 557let Defs = [CC], CCValues = 0xC in { 558 def TCEB : TestRXE<"tceb", 0xED10, z_tdc, FP32>; 559 def TCDB : TestRXE<"tcdb", 0xED11, z_tdc, FP64>; 560 def TCXB : TestRXE<"tcxb", 0xED12, z_tdc, FP128>; 561} 562 563//===----------------------------------------------------------------------===// 564// Floating-point control register instructions 565//===----------------------------------------------------------------------===// 566 567let hasSideEffects = 1 in { 568 let mayLoad = 1, mayStore = 1 in { 569 // TODO: EFPC and SFPC do not touch memory at all 570 let Uses = [FPC] in { 571 def EFPC : InherentRRE<"efpc", 0xB38C, GR32, int_s390_efpc>; 572 def STFPC : StoreInherentS<"stfpc", 0xB29C, storei<int_s390_efpc>, 4>; 573 } 574 575 let Defs = [FPC] in { 576 def SFPC : SideEffectUnaryRRE<"sfpc", 0xB384, GR32, int_s390_sfpc>; 577 def LFPC : SideEffectUnaryS<"lfpc", 0xB29D, loadu<int_s390_sfpc>, 4>; 578 } 579 } 580 581 let Defs = [FPC], mayRaiseFPException = 1 in { 582 def SFASR : SideEffectUnaryRRE<"sfasr", 0xB385, GR32, null_frag>; 583 def LFAS : SideEffectUnaryS<"lfas", 0xB2BD, null_frag, 4>; 584 } 585 586 let Uses = [FPC], Defs = [FPC] in { 587 def SRNMB : SideEffectAddressS<"srnmb", 0xB2B8, null_frag, shift12only>, 588 Requires<[FeatureFPExtension]>; 589 def SRNM : SideEffectAddressS<"srnm", 0xB299, null_frag, shift12only>; 590 def SRNMT : SideEffectAddressS<"srnmt", 0xB2B9, null_frag, shift12only>; 591 } 592} 593 594//===----------------------------------------------------------------------===// 595// Peepholes 596//===----------------------------------------------------------------------===// 597 598def : Pat<(f32 fpimmneg0), (LCDFR_32 (LZER))>; 599def : Pat<(f64 fpimmneg0), (LCDFR (LZDR))>; 600def : Pat<(f128 fpimmneg0), (LCXBR (LZXR))>; 601